debate

Debate — Multi-Agent Adversarial Reasoning

Spawn 2-3 AI agents with competing perspectives on a question, collect their arguments, and synthesize the strongest conclusion.

Step graph (mirrors frontmatter steps:): step-1-parse-input → step-2-build-prompts → step-3-dispatch-agents → step-4-synthesize → step-5-optional-round-2

Critical

• Synthesis (step-4) is where the value is — don't just summarize. Read every agent's output, identify genuine tensions, name what they agree on. The orchestrator's authorship of the synthesis IS the skill; if all you do is concatenate, you wasted the cost.
• Gaps MUST be named when an agent fails. .gap files written by dispatch-agents.sh mean an agent never returned; the Gaps section MUST surface "Agent X failed (timeout/rate-limit). Synthesis reflects N of M — the {role} position may be underrepresented." Do not silently emit a 2-agent synthesis labeled 3-agent.
• Stop when a round adds nothing. Per @rule:adversarial-recursion-cap, the loop ends on the work: a round that surfaces no argument already on the table is the last one. There is no round number to hit and no question to fire — rounds here are requested by the user in the first place.

step-1-parse-input

Extract the question from $ARGUMENTS. Parse optional flags:

• --format: dialectic (default), redteam, council
• --agents: 2 (default) or 3 — how many seats the debate has. WHO fills them is decided at dispatch by which model CLIs you have installed, not here.

Auto-detect format if not specified:

• Binary choice / "X or Y" / "X vs Y" → dialectic
• Evaluating a plan/proposal/design → redteam
• Open-ended, multi-dimensional → council

If the question is unclear or too vague, ask ONE clarifying question first.

Context Scaffold

Before building prompts, assemble a context block. If the user gave a bare question, check whether it references a known project/file/decision and pull key facts. Structure:

CONTEXT FOR DEBATE:
- Decision: [one sentence framing the choice]
- Constraints: [timeline, budget, dependencies, blockers]
- Stakes: [what's at risk if the wrong choice is made]
- Current leaning: [user's position, if any — agents should challenge this]